Kamal Ghorui
When software engineer Steve Yegge published a lengthy thread claiming Google‘s AI adoption was roughly on par with John Deere—yes, the tractor company—the response from the top was swift and blunt. Demis Hassabis, CEO of Google DeepMind, told Yegge to have his source “do some actual work and stop spreading absolute nonsense,” calling the post “completely false and just pure clickbait.“Yegge’s original thread, which went viral over the weekend, was based on a conversation with a longtime Google tech director. The core claim: Google engineers couldn’t lean into agentic coding the way the rest of the industry was because using Claude Code—which Yegge described as the current benchmark—was seen internally as using the enemy’s product. Gemini, he argued, had never been compelling enough to replace it. The result, according to Yegge, was a company coasting in mediocrity while a hiring freeze kept anyone from realizing how far behind they’d fallen.
Google says 40,000 engineers are already doing agentic coding every week
Yegge didn’t fully back down. He challenged Google to prove that half its engineers were burning at least 4 million tokens a day—a threshold he said would earn a public retraction—and noted that at Anthropic, that figure runs somewhere between 10 and 15 million tokens per engineer daily.Google engineering director Addy Osmani pushed back with more specifics. He said over 40,000 Google engineers use agentic coding tools on a weekly basis and have access to the company’s own CLI tools, custom models, MCP integrations, orchestrators, and virtual SWE team setups. He also noted that many Googlers actively use external tools—including Claude Code—through Anthropic’s models on Vertex, partly to benchmark their own work.
Inside the broader AI adoption gap splitting the tech industry in two
Yegge’s broader argument—that a prolonged hiring freeze had siloed companies from each other, leaving most flying blind on where they stood in the AI adoption race—still stands uncontested. He described a familiar curve across the industry: roughly 20% power users going fully agentic, 20% outright refusing to adopt AI tools, and 60% still using basic chat interfaces like Cursor without going deeper.
